Scott Walters > Math-Preference-SVD > Math::Preference::SVD

Download:
Math-Preference-SVD-0.01.tar.gz

Dependencies

Annotate this POD

View/Report Bugs
Module Version: 0.01   Source  

NAME ^

Math::Preference::SVD - Preference/Recommendation Engine based on Single Value Decomposition

SYNOPSIS ^

    use Math::Preference::SVD;

    my $x = Math::Preference::SVD->new;

    my @users = (0..3);
    my @movies = (0..3);

    my @ratings = (
        map({ [ $_, 0, 4 ] } @users),    # *everyone* says item 0 is rated 4
        map({ [ $_, 1, 5 ] } @users),    # *everyone* says item 1 is rated 5
        map({ [ $_, 2, 1 ] } @users),    # *everyone* says item 2 is rated 1
        map({ [ $_, 3, 2 ] } @users),    # *everyone* says item 3 is rated 2
    );

    $x->set_ratings( \@ratings, );

    for my $cust (@users) {
        for my $movie (@movies) {
            # predict_rating() takes movie_id then cust_id -- yes, 
            # this seems backward to me too
            my $predicted = sprintf "%1.2f", $x->predict_rating($movie, $cust);
            print "cust $cust says about movie $movie: predicted: $predicted";
        }
    }

DESCRIPTION ^

This module imples a simple "preference engine" based on one of the entries to the NetFlix Prize competition. Preference engines take user rating data for items and attempt to predict the user's rating for other items so that a system might find and suggest to them other things they're likely to purchase or enjoy.

Single Value Decoposition takes a large rectangular array of data and decomposes it into two matrices, one as long as the data and one as wide, that approximates the original matrix. here and there.

And then it does it again, starting with the error left off by the first set of matrices, making a second set of long and wide matrices. And then it does it again. The result is a series of matrices that can be multiplied together, and their outputs totaled up, to approximately reconstruct the original.

The large input matrix might, for instance, have customers in the columns and movies in the rows, with data filled in here and there to specify that customer's rating of that movie. Each set of matrices, with a tall one (for movies) and a wide one (for customers), could be thought of as containing information about some tangible attribute of the movies being rated and how the user feels about that attribute. See the references below for an awesome example of the first three attributes extracted from the NetFlix Prize data.

Extrapolation is a side-effect of this lossy compression scheme. This is where the recommendation engine bit comes into play -- you can ask the thing for ratings by a customer that that customer never made, and it'll dutifully multiply the customer's features versus the movie's features for the various features, add them up, and give you a predicted rating for that movie for that user. By iterating over all of the movies (and I say movies, but the user could be rating anything, including other users) and asking for predictions for a specific user for that movie, and sorting the results, you can come up with a list of recommendations that he or she or it might like.

Deep, Profound Suckage

This is Alpha software! Input validation is poor or non-existant, the C code likes to coredump, and there's presently no way to adjust tuning parameters without editing the XS, and the thing really needs tuning to the dataset at hand.

For what it's worth, SVD is nice for computing lots of recommendations quickly, but due to the extremely lossy nature, it really isn't a fantastic recommendation engine. It also requires large amounts of tuning to work on any data set. Quoting from http://sifter.org/~simon/journal/20061211.html on the subject of the #defines inside the C:

"Despite the regularization term in the final incremental law above, over fitting remains a problem. Plotting the progress over time, the probe rmse eventually turns upward and starts getting worse (even though the training error is still inching down). We found that simply choosing a fixed number of training epochs appropriate to the learning rate and regularization constant resulted in the best overall performance. I think for the numbers mentioned above it was about 120 epochs per feature, at which point the feature was considered done and we moved on to the next before it started over fitting."

See the URL for a more complete description of the over-fitting problem, but the short version is that if it tries too hard at first by iterating too many times in the successive approximation feedback thingie, then the first feature (set of wide and tall matrices) will fit well, but all of the ones after it will fit worse, and the overall quality of predictions will drop.

API

Nothing is exported or available for export. Use the OO interface -- even though you're allowed only one object because the C allocates one set of datastructures, once.

    use Math::Preference::SVD;
    my $x = Math::Preference::SVD->new;

new() takes no parameters.

        map({ [ $_, 3, 2 ] } @users),    # *everyone* says item 3 is rated 2
    );

    $x->set_ratings( [ [ cust_id_1, movie_id_1, rating ], ..., ] );

set_ratings() takes all of the data in one batch. It takes an arrayref full of arrayrefs, each of those containing three fields: the customer id of the person doing the rating; the numeric id of the thing being rated, and the rating itself. The rating must be 1-5 and this is enforced inside the engine.

Taking all of the data in one batch done by the Perl API so that it can extract the largest customer_id and movie_id so that the C datastructures can be correctly sized. See below for instructions on how to bang the XS directly if you want to optimize by incrementally loading the data and are willing to size these structures yourself.

    $x->predict_rating($movie, $cust);

Yes, this is ass backwards from set_ratings(). Sorry. I'll fix it in the next version. Gets a floating point value of the range 1.0 through 5.0 inclusive.

Incrementally Loading Rating Data

If you want to pre-size the datastructures and then load data incrementally, ditch the OO interface and call the XS routines directly using fully qualified package names. The Engine() function takes the sizes of the data structures, then set_Movies() is called for each data item to be rated then in a nested loop, set_Ratings() is called. Movies (or whatever -- I'm just going to continue to call them movies) must be loaded in order with all of their ratings loaded sometime after them and perhaps immediately after them -- I haven't tested. See the Perl source of this module for an example of this usage -- it's pretty brief.

XS API

  void Engine(int, int, int);
  void set_Movies(int, int, int);    // set_Movies($movieId, 0, 0); # RatingCount = 0, RatingSum = 0
  void set_Ratings(int, int, int);   // set_Ratings($movieId, $custId, $rating);
  void CalcMetrics();                // call after loading all of the rating data -- pre-calc
  void CalcFeatures();               // call after calling the above -- does the actual work
  double PredictRating(short, int);  // takes $movieId, $custId, returns predicted rating, 1-5 float
  void DestroyEngine();              // free up the working RAM -- untested

EXPORT

None.

SEE ALSO ^

As far as Perl modules: Dunno. Any suggestions for me?

http://www.timelydevelopment.com/demos/NetflixPrize.aspx
http://www.kdnuggets.com/news/2007/n08/6i.html

BUGS ^

Values out of range make it coredump. Other things make it coredump. Duplicate ratings (same user and movie) seem to make it coredump. You can only create one object at a time. Memory is not automatically freed but must be manually freed. The algorithm is kind of lame and needs lots of tuning. The "other users with similar tastes" thing is a joke.

AUTHOR ^

Hi, I'm Scott, and I did this Perl adaption and wrote this documentation, and this is my copyright foreword: This code is covered by _two_ copyrights; I have copyright over my own work, and Timely Development, LLC has copyright over their work. As such, to copy this entire thing, you must adhere to both copyrights. Among other things, this means no stripping off credits or copyright notices just because GPL+Artistic touched this thing. Fact of the matter is, the vast majority of the code is Timely Development, LLC's, not mine. Thanks. I'm glad we could have this little chat.

Copyright (c) 2008 by Scott Walters

This library is free software; you can redistribute it and/or modify it under the same terms as Perl itself, either Perl version 5.8.8 or, at your option, any later version of Perl 5 you may have available.

Copyright (c) 2008 by Timely Development, LLC.

# SVD Sample Code # # Copyright (C) 2007 Timely Development (www.timelydevelopment.com) # # Special thanks to Simon Funk and others from the Netflix Prize contest # for providing pseudo-code and tuning hints. # # Feel free to use this code as you wish as long as you include # these notices and attribution. # # Also, if you have alternative types of algorithms for accomplishing # the same goal and would like to contribute, please share them as well :) # # STANDARD DISCLAIMER: # # - THIS CODE AND INFORMATION IS PROVIDED "AS IS" WITHOUT WARRANTY # - OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING BUT NOT # - LIMITED TO THE IMPLIED WARRANTIES OF MERCHANTABILITY AND#OR # - FITNESS FOR A PARTICULAR PURPOSE. #

syntax highlighting: